ThinkRouter

---

The problem

Reasoning models (o1, DeepSeek-R1, Claude thinking) apply the same 8,000-token compute budget to every query — whether it is simple arithmetic or a complex proof.

"What is 2 + 3?"                   →  8,000 thinking tokens   ← 99% wasted
"Prove that sqrt(2) is irrational"  →  8,000 thinking tokens   ← correctly used

At 100,000 queries per day, that is $192,635/month in avoidable spend.

---

The solution

from thinkrouter import ThinkRouter

client   = ThinkRouter(provider="openai")
response = client.chat("What is the capital of France?")
# Routed to NO_THINK → 50 tokens used, not 8,000

client.usage.print_dashboard()

  ThinkRouter — Usage Dashboard
  ──────────────────────────────────────────────
  Total calls          : 13
  Tokens saved         : 55,650
  Compute savings      : 53.5%
  Avg classifier time  : 0.02 ms

  Routing breakdown:
    no_think        :      7  (53.8%)  — Direct answer
    short_think     :      0  ( 0.0%)  — Moderate reasoning
    full_think      :      6  (46.2%)  — Full extended reasoning

---

How it works

ThinkRouter intercepts each query, runs a lightweight classifier in under 1ms, and routes to the minimum compute budget:

Tier	Budget	Use case
NO_THINK	50 tokens	Arithmetic, definitions, lookups, translations
SHORT	800 tokens	Multi-step reasoning, moderate chaining
FULL	8,000 tokens	Proofs, system design, algorithm implementation

---

Installation

# Base install — works immediately, zero ML dependencies
pip install thinkrouter

# With fine-tuned DistilBERT classifier (higher accuracy)
pip install thinkrouter[classifier]

# With OpenAI client
pip install thinkrouter[openai]

# With Anthropic client
pip install thinkrouter[anthropic]

# Everything
pip install thinkrouter[all]

---

Quick start

Try it now — no API key needed

OpenAI

from thinkrouter import ThinkRouter

client = ThinkRouter(
    provider="openai",
    api_key="sk-...",      # or set OPENAI_API_KEY
    model="gpt-4o",
    verbose=True,
)

response = client.chat("Explain how merge sort works.")
print(response.content)
print(response.routing)
# ClassifierResult(tier=FULL, confidence=0.87, budget=8000 tokens, latency=1.2ms)

client.usage.print_dashboard()

Anthropic

client = ThinkRouter(
    provider="anthropic",
    api_key="sk-ant-...",  # or set ANTHROPIC_API_KEY
    model="claude-haiku-4-5-20251001",
)

response = client.chat("What is 144 divided by 12?")
# Routed to NO_THINK → 50 tokens, not 8,000

Streaming

for chunk in client.stream("Explain quantum entanglement step by step."):
    print(chunk, end="", flush=True)

Classify without an API call

results = client.classify_batch([
    "What is 7 * 8?",
    "Design a distributed caching system.",
    "How many days are in a leap year?",
])

for r in results:
    print(f"{r.tier.name:<12}  budget={r.token_budget:>6} tokens  conf={r.confidence:.2f}")

NO_THINK      budget=    50 tokens  conf=0.88
FULL          budget=  8000 tokens  conf=0.85
NO_THINK      budget=    50 tokens  conf=0.80

---

Cost savings at scale

Volume	Savings/day	Savings/month
10,000 queries/day	$642	$19,263
100,000 queries/day	$6,421	$192,635
1,000,000 queries/day	$64,212	$1,926,346

Based on 53.5% savings rate, $15/million reasoning tokens (approximate o1 rate).

---

Classifier backends

Heuristic (default)

Zero dependencies. Regex patterns and word-count heuristics. Runs in under 1ms.

client = ThinkRouter(classifier_backend="heuristic")

DistilBERT

Fine-tuned on GSM8K. Achieves 93%+ quality retention at 60% compute savings.
Requires pip install thinkrouter[classifier].

client = ThinkRouter(
    classifier_backend="distilbert",
    confidence_threshold=0.75,
)

---

Confidence threshold

Threshold	Savings	Quality retained	Use case
0.65	~59%	~91%	High cost sensitivity
0.75	~55%	~93%	Recommended
0.85	~44%	~96%	Quality-sensitive

Queries below the threshold fall back to FULL — never degrades output quality.

---

API reference

ThinkRouter

ThinkRouter(
    provider             = "openai",      # "openai" | "anthropic" | "generic"
    api_key              = None,          # falls back to OPENAI_API_KEY / ANTHROPIC_API_KEY
    model                = None,          # default model for all calls
    classifier_backend   = "heuristic",   # "heuristic" | "distilbert"
    confidence_threshold = 0.75,
    max_records          = 10_000,
    verbose              = False,
)

RouterResponse

response.content       # str — generated text
response.routing       # ClassifierResult
response.provider      # "openai" | "anthropic"
response.model         # model identifier
response.usage_tokens  # {"prompt_tokens": N, "completion_tokens": M, ...}

ClassifierResult

result.tier          # Tier.NO_THINK | Tier.SHORT | Tier.FULL
result.confidence    # float in [0, 1]
result.token_budget  # int — thinking tokens assigned
result.latency_ms    # classifier wall-clock time in ms
result.backend       # "heuristic" | "distilbert:cuda" | "distilbert:cpu"

---

Running tests

git clone https://github.com/saikoushiknalubola/thinkrouter.git
cd thinkrouter
pip install -e ".[dev]"
pytest tests/ -v

---

Roadmap

• Heuristic classifier
• OpenAI and Anthropic adapters
• Streaming support
• Thread-safe usage dashboard
• GitHub Actions CI (Python 3.9–3.12)
• DistilBERT model on HuggingFace Hub
• Multi-domain training (MMLU, HumanEval, ARC-Challenge)
• Async support (achat(), astream())
• Continuous budget regression
• Hosted API proxy (api.thinkrouter.ai)

---

Research basis

• Zhao et al. (2025). SelfBudgeter. arXiv:2505.11274 — 74.47% savings validated
• Wang et al. (2025). TALE-EP. ACL Findings 2025 — 67% output token reduction
• Sanh et al. (2019). DistilBERT. arXiv:1910.01108
• Cobbe et al. (2021). GSM8K. arXiv:2110.14168

---

Contributing

See CONTRIBUTING.md. Issues and pull requests welcome.

---

License

MIT — see LICENSE.